Description
Describe the feature request
I've been trying to convert the Qwen3 0.6B Embedding Model into ONNX Format using optimum-cli and this error was thrown saying it is not supported yet.
optimum-cli export onnx --library transformers --task sentence-similarity -m Qwen/Qwen3-Embedding-0.6B --optimize O3 model-output-dir
NotImplementedError: ONNX Runtime doesn't support the graph optimization of qwen3 yet. Only ['albert', 'bart', 'bert', 'big-bird', 'bigbird-pegasus', 'blenderbot', 'bloom', 'camembert', 'codegen', 'deberta', 'deberta-v2', 'dinov2', 'distilbert', 'electra', 'gpt2', 'gpt-bigcode', 'gpt-neo', 'gpt-neox', 'gptj', 'granite', 'longt5', 'llama', 'marian', 'mbart', 'mistral', 'modernbert', 'mpnet', 'mt5', 'm2m-100', 'nystromformer', 'pegasus', 'roberta', 'segformer', 't5', 'vit', 'whisper', 'xlm-roberta', 'pix2struct'] are supported. If you want to support qwen3 please propose a PR or open up an issue in ONNX Runtime: https://github.com/microsoft/onnxruntime.
Describe scenario use case
Qwen3 0.6B embeddings is a really good embedding model ranking 4th at the MTEB leaderboard. It will be a great update for the retrieval systems that uses open source embedding models.